[image 00346] 11/13(水)チュートリアル開催のご案内

[Akihiro Sugimoto]

みなさま

下記の予定でパターン認識に関するチュートリアルを開催します
ので、ふるってご参加ください。事前登録不要、参加費無料です。

杉本
------------------
日時:  2013年11月13日(水) 13:30〜
場所:  国立情報学研究所　20F 2004号室
　　　(東京都千代田区一ツ橋)
            http://www.nii.ac.jp/about/access/
講師: 
   Vaclav Hlavac
    Professor
    Czech Technical University in Prague, Czech Republic
   http://cmp.felk.cvut.cz/~hlavac

Title: The Appetizer to Pattern Recognition, a tutorial

Introduction:

Being the visiting professor at NII, we agreed with my NII host Prof.
Akihiro Sugimoto that it might be plausible to give a tutorial to basics of
pattern recognition (called also machine learning). The tutorial should
consist of two or (optionally three) lectures 60 minutes long. No special
foreknowledge is expected. The topics should be useful practically. The
tutorial will be given in a single afternoon to enable the audience to use
their time efficiently.

Tutorial 1: Informal introduction to pattern recognition

The core of pattern recognition (machine learning) is to learn a decision
rule (a classifier) from the empirical experience, i.e. from observed
examples. This approach will be motivated and put into a wider context. The
need for statistical approach will be explained. The classifier will be
introduced first informally. The Bayesian formulation unifies nicely various
approaches. Based on it, it will be explained what tasks can be fulfilled in
pattern recognition and which not. This means that the pattern recognition
territory will be outlined in the tutorial informally.

Tutorial 2: Evaluation of the classifier performance

Classifiers (decision rules) learned often empirically from examples either
in a supervised manner from labeled examples provided by a "teacher" or in
an unsupervised manner. The natural question arises: Can the classifier
performance be evaluated empirically? Is the classifier able to generalize?
Will it perform well on data unseen during its learning? Is it accurate
enough? There is an established methodology stemming in statistical
hypotheses testing used in this context. It will be explained together with
concepts like confusion matrix and ROC (Receiver Operating Characteristic).

Tutorial 3: Learning formulated as an optimization task, several optimality
criteria 
(optional, only if the audience of previous two tutorials wants it)

Learning has facets of dreaming in a fairy tale manner as well as the
rigorous mathematical formulation as an optimization task. The tutorial will
discuss relation between these two limiting standpoints. Statistical
learning will be represented in Bayesian framework, in which the optimally
learned classifier minimizes the Bayesian risk. The trouble is that
empirical data provided by a training set does not provide the needed
information. Four substitutive optimization criteria will be introduced and
discussed. The approach provides a plausible insight into classifier
learning.

----------